YesPress / Company File №57
Fundamental.
San Francisco's bet that the next great foundation model will not write your novel - it will close your books, forecast your demand, and explain your churn.
San Francisco's bet that the next great foundation model will not write your novel - it will close your books, forecast your demand, and explain your churn.
It is May 2026, and somewhere on an AWS dashboard, a Fortune 100 supply chain planner is buying a foundation model the way she buys storage. She picks a region. She picks an instance. She picks NEXUS. By the end of the afternoon, a model that has never been shown a single sentence of English is forecasting next quarter's inventory - and getting it more right than the team's twelve-year-old XGBoost pipeline. The model belongs to Fundamental, a 57-person company in San Francisco that raised $255 million on the unfashionable idea that the most valuable data in the world is not on the internet. It is in a database somewhere, in rows and columns, waiting.
The AI boom has been a triumph of unstructured data. Text, images, audio, video - every modality with a creative-writing teacher behind it got a foundation model. Tables, the modality that actually runs payroll, supply chains, retail pricing, insurance, and most of the Fortune 500, got more pipelines. More feature engineering. More dashboards. The kind of work that nobody puts on a keynote slide.
Fundamental's founders looked at this and saw something Wilde would have appreciated: the world had invested billions of dollars and a generation of PhDs in teaching machines to talk, while the data that signs the checks was still being prepped by hand. There is an irony there, and Fundamental decided it was an opportunity.
Jeremy Fraenkel does not look like a guy who needed another company. He had already done two startups and exited both. Before that he had been at JP Morgan and Bridgewater, which is to say he had spent his early career inside the institutions that, more than anyone, live and die by structured data. He took a graduate degree in machine learning at Berkeley and then, in October 2024, did something that on paper looked like a mistake: he started a research company devoted to a category that did not exist yet.
He called the category Large Tabular Models. The acronym, LTM, is a deliberate echo of LLM - and a quiet provocation. If you can pre-train a model on the entire public internet and have it write sonnets, the argument goes, you can pre-train one on billions of real enterprise tables and have it predict revenue. The investors agreed. A $30 million seed in 2024 turned into a $225 million Series A on February 6, 2026, led by Oak HC/FT with Salesforce Ventures, Valor Equity Partners, Battery Ventures, and Hetz Ventures climbing in behind them.
NEXUS is the company's flagship and, at the moment, its only public product. It is a foundation model, in the strict technical sense: pre-trained once, on billions of real-world enterprise tables, then deployed against any new table without bespoke modeling. Demand forecasting. Price prediction. Churn. The model does not need to be told what the columns mean. It works it out.
The pitch to engineers is almost rude in its simplicity. Connect your data. Call the model. Skip the feature engineering, skip the model design, skip the six weeks of arguing with the data team. Treat business data, in Fundamental's words, "not as a simple sequence of words, but as a complex web of non-linear relationships." Which is what databases have been all along - and what nobody, until now, was willing to train a foundation model on at scale.
Source: Fundamental announcement, Feb 2026 - Salesforce Ventures, Valor, Battery, Hetz round out the cap table.
Above: a chart you could have drawn on a napkin, which is the point. The interesting question is not how much they raised. It is who paid attention.
For a company that emerged from stealth fewer than four months ago, Fundamental has been busy. The public references are still thin - Fortune 100 customers tend to dislike press releases - but the company has confirmed seven-figure contracts in production for demand forecasting, price prediction, and customer churn. The partners list reads like a credibility checklist for enterprise AI: Amazon Web Services for distribution, SAP for the data plane that most of the global enterprise actually lives on. AWS is the headline. Selling a foundation model the way you sell a virtual machine is not the kind of partnership a hyperscaler grants on a whim.
"The Power to Predict" is the four-word tagline on fundamental.tech, and it is doing a lot of work. The full mission, as the company describes it, is to give enterprises a foundation model native to the structured data that actually runs their business - not adapted from text, not bolted onto a chat interface, not pretending to think when it should just be predicting. The vision is bigger and quieter: a world where every business decision is informed by a model that understands the company's data the way an LLM understands language.
It is, in its way, an anti-glamour mission. LTMs will not write your wedding speech. They will tell you, with uncomfortable accuracy, how many units you will sell next Tuesday in the Pacific Northwest, and why. Fundamental seems comfortable with that trade.
There is a version of the next five years in which LLMs and LTMs sit together inside every meaningful enterprise stack - one for language and reasoning, one for the numbers that determine whether the lights stay on. In that version, Fundamental's bet looks obvious in hindsight, the way Snowflake's data-warehouse-as-a-service bet looks obvious now, and the way nobody admits they were skeptical in 2014. There is also a version where LTMs remain a niche, classical ML quietly improves, and the lakehouse wins on inertia alone. Fundamental's strategy - foundation-model rigor, hyperscaler distribution, Fortune 100 anchor customers - is built for the first version and survivable in the second.
Either way, the conversation in enterprise AI has shifted. A year ago, "tabular foundation model" was a phrase you used to win an argument at a research conference. Today, it is a line item on an AWS invoice.
It is May, still. The supply chain planner closes the laptop. The forecast that took six weeks last year took her an afternoon. She did not write a feature pipeline. She did not argue with the data team. She picked a model in a console and pointed it at a table. The model has never read a sentence in its life. It read her data instead, and that turned out to be enough. That is what Fundamental built, in eighteen months, with $255 million and 57 people who decided the unglamorous half of the enterprise was the half worth a foundation model.
The interesting models, it turns out, are not always the ones with anything to say.